Security is a priority task of the current passenger terminals. In order to meet the actual security requirements of explosives in passenger terminals, this paper uses a CO2 laser to generate a uniform laser, which is expended by an expansion system, and then to irradiate the object under test. This paper uses the high resolution infrared thermal imager to collect the laser irradiation area, and the temperature image of the measured object is extracted by software, the neural network algorithm is combined to check and judge the temperature image. In this paper, an improved deep full convolutional neural network Detectnet is proposed to detect explosives in thermal images. The network consists of extended convolutional layer, backbone network and spatial pyramid pool. These modules are conducive to efficient feature extraction of explosives, image aggregation and resolution reconstruction, and they can greatly improve the neural network's extraction function of micro-trace explosives in detected objects. Compared with traditional image processing to detect explosives, the network proposed in this paper has higher efficiency and faster detection speed, and can automatically distinguish the size and location of TNT and RDX contaminated areas. The accuracy rate of the network proposed in this paper is 146% of that of traditional method, while the running time of the network proposed in this paper is about 12% of that of traditional method for detecting the same image. The network proposed in this paper can make the detection and identification of explosives in security check tasks more accurate and faster.